RISC-V behavior tests disabled

[andrewrk]

These are all upstream issues with LLVM.

@newStackCall causes:

LLVM ERROR: Named registers not implemented for this target

Issues with f16:

LLVM ERROR: Cannot select: 0x39d32ee0: f32,ch = load<(dereferenceable load 2 from %ir.a), anyext from f16> 0x39d29678, FrameIndex:i64<0>, undef:i64, widening.zig:23:18
  0x39d1aa68: i64 = FrameIndex<0>
  0x38d69710: i64 = undef
In function: behavior.widening.test "float widening"

LLVM ERROR: Cannot select: 0x393500f8: f32 = fp16_to_fp Constant:i64<18988>, widening.zig:40:19
  0x39350298: i64 = Constant<18988>
In function: behavior.widening.test "float widening f16 to f128"

[luismarques]

Hi. I'm one of the LLVM RISC-V target developers. I tried to diagnose this issue but I can't seem to figure out how to crosscompile the zig test suite. The -target option doesn't seem to work for the test command, and the documentation is scarce.

BTW, I suggest also testing zig releases with debug builds of LLVM. The zig 0.5 release seems to trip a few LLVM asserts, I had to comment-out one of them to get the zig test suite to compile (for the x86 host).

[andrewrk]

Welcome, @luismarques.

The -target option doesn't seem to work for the test command

This functionality has test coverage. What did you try, and what output did you get?

BTW, I suggest also testing zig releases with debug builds of LLVM. The zig 0.5 release seems to trip a few LLVM asserts, I had to comment-out one of them to get the zig test suite to compile (for the x86 host).

I did test this, with the merge of the llvm9 branch. If Zig is tripping any LLVM asserts, that is a regression since e81156b.

documentation is scarce

Have a look at https://github.com/ziglang/zig/blob/master/CONTRIBUTING.md. Ideally we would be able to do this to RISC-V test coverage:

--- a/test/tests.zig
+++ b/test/tests.zig
@@ -70,6 +70,36 @@ const test_targets = [_]TestTarget{
         .link_libc = true,
     },
 
+    TestTarget{
+        .target = Target{
+            .Cross = CrossTarget{
+                .os = .linux,
+                .arch = builtin.Arch.riscv64,
+                .abi = .none,
+            },
+        },
+    },
+    TestTarget{
+        .target = Target{
+            .Cross = CrossTarget{
+                .os = .linux,
+                .arch = builtin.Arch.riscv64,
+                .abi = .musl,
+            },
+        },
+        .link_libc = true,
+    },
+    TestTarget{
+        .target = Target{
+            .Cross = CrossTarget{
+                .os = .linux,
+                .arch = builtin.Arch.riscv64,
+                .abi = .gnu,
+            },
+        },
+        .link_libc = true,
+    },
+
     TestTarget{
         .target = Target{
             .Cross = CrossTarget{

If you apply this patch and then:

./zig build test -Denable-qemu

This will reveal RISC-V problems, when you re-enable the tests that are disabled. You can find these by grepping for riscv (from the build directory):

$ grep -RI riscv ../test/
../test/stage1/behavior/new_stack_call.zig:    if (@import("builtin").arch == .riscv64) {
../test/stage1/behavior/math.zig:    if (@import("builtin").arch == .riscv64) {
../test/stage1/behavior/math.zig:    if (@import("builtin").arch == .riscv64) {
../test/stage1/behavior/math.zig:    if (@import("builtin").arch == .riscv64) {
../test/stage1/behavior/cast.zig:    if (@import("builtin").arch == .riscv64) {
../test/stage1/behavior/widening.zig:    if (@import("builtin").arch == .riscv64) {
../test/stage1/behavior/widening.zig:    if (@import("builtin").arch == .riscv64) {

I'm looking forward to working with you. I'm excited for an open hardware standard such as RISC-V.

[luismarques]

This functionality has test coverage. What did you try, and what output did you get?

I tried many combinations, I no longer recall all of them. Here's my reasoning. This works for the host (kind of, it ends up tripping an assertion):

zig-0.5.0/build$ bin/zig build test
(...)
768/768 behavior.widening.test "behavior-x86_64-linux-musl-Debug-c-multi float widening f16 to f128"...OK
All tests passed.
zig: ../lib/CodeGen/GlobalISel/MachineIRBuilder.cpp:907: void llvm::MachineIRBuilder::validateSelectOp(const llvm::LLT &, const llvm::LLT &, const llvm::LLT &, const llvm::LLT &): Assertion `(ResTy == Op0Ty && ResTy == Op1Ty) && "type mismatch"' failed.

...and --help says this:

Compile Options:
(...)
  -target [name]               <arch><sub>-<os>-<abi> see the targets command

Therefore I would expect something like this to work:

$ bin/zig build -target riscv64-linux test
Unrecognized argument: -target
$ bin/zig -target riscv64-linux build test
Unrecognized command: build
$ bin/zig build test -target riscv64-linux
Unrecognized argument: -target

I did test this, with the merge of the llvm9 branch. If Zig is tripping any LLVM asserts, that is a regression since e81156b.

You can download http://releases.llvm.org/9.0.0/llvm-9.0.0.src.tar.xz and http://releases.llvm.org/9.0.0/cfe-9.0.0.src.tar.xz and reproduce this issue. I used the following llvm/cfe cmake configurations:

cmake -G Ninja -DCMAKE_BUILD_TYPE="Debug" \
  -DBUILD_SHARED_LIBS=True -DLLVM_USE_SPLIT_DWARF=True \
  -DLLVM_BUILD_TESTS=True \
  -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ \
  -DLLVM_ENABLE_LLD=True \
  -DLLVM_TARGETS_TO_BUILD="all" \
  -DLLVM_APPEND_VC_REV=False \
  -DLLVM_ENABLE_LIBXML2=OFF \
  -DLLVM_EXPERIMENTAL_TARGETS_TO_BUILD="AVR" \
  -DCMAKE_INSTALL_PREFIX=$HOME/tools/llvm-9 -DCMAKE_PREFIX_PATH=$HOME/tools/llvm-9 \
  ..

cmake  -G Ninja -DCMAKE_BUILD_TYPE="Debug" \
  -DBUILD_SHARED_LIBS=True -DLLVM_USE_SPLIT_DWARF=True \
  -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ \
  -DCMAKE_INSTALL_PREFIX=$HOME/tools/llvm-9 -DCMAKE_PREFIX_PATH=$HOME/tools/llvm-9 \
  ..

Have a look at https://github.com/ziglang/zig/blob/master/CONTRIBUTING.md. Ideally we would be able to do this to RISC-V test coverage:

I will continue to take a look at this, as time permits. Thanks.

I'm looking forward to working with you. I'm excited for an open hardware standard such as RISC-V.

Likewise!

[andrewrk]

I fixed the LLVM assertion failure in a8d6954. I'm running the full test suite locally with a debug build of LLVM to find if there were any other regressions. (edit: found one: #3358)

bin/zig build tests all the targets, not just one. Here's the most direct way to reproduce this issue:

1. Apply the following patch:

diff --git a/test/stage1/behavior/cast.zig b/test/stage1/behavior/cast.zig
index 8176f5cf..04c7fa60 100644
--- a/test/stage1/behavior/cast.zig
+++ b/test/stage1/behavior/cast.zig
@@ -247,10 +247,6 @@ fn testPeerErrorAndArray2(x: u8) anyerror![]const u8 {
 }
 
 test "@floatToInt" {
-    if (@import("builtin").arch == .riscv64) {
-        // TODO: https://github.com/ziglang/zig/issues/3338
-        return error.SkipZigTest;
-    }
     testFloatToInts();
     comptime testFloatToInts();
 }
diff --git a/test/stage1/behavior/math.zig b/test/stage1/behavior/math.zig
index 1ff63e3e..e285d0eb 100644
--- a/test/stage1/behavior/math.zig
+++ b/test/stage1/behavior/math.zig
@@ -6,10 +6,6 @@ const maxInt = std.math.maxInt;
 const minInt = std.math.minInt;
 
 test "division" {
-    if (@import("builtin").arch == .riscv64) {
-        // TODO: https://github.com/ziglang/zig/issues/3338
-        return error.SkipZigTest;
-    }
     testDivision();
     comptime testDivision();
 }
@@ -573,10 +569,6 @@ fn remdiv(comptime T: type) void {
 }
 
 test "@sqrt" {
-    if (@import("builtin").arch == .riscv64) {
-        // TODO: https://github.com/ziglang/zig/issues/3338
-        return error.SkipZigTest;
-    }
     testSqrt(f64, 12.0);
     comptime testSqrt(f64, 12.0);
     testSqrt(f32, 13.0);
@@ -622,10 +614,6 @@ test "vector integer addition" {
 }
 
 test "NaN comparison" {
-    if (@import("builtin").arch == .riscv64) {
-        // TODO: https://github.com/ziglang/zig/issues/3338
-        return error.SkipZigTest;
-    }
     testNanEqNan(f16);
     testNanEqNan(f32);
     testNanEqNan(f64);
diff --git a/test/stage1/behavior/new_stack_call.zig b/test/stage1/behavior/new_stack_call.zig
index a2f3773a..2fef4816 100644
--- a/test/stage1/behavior/new_stack_call.zig
+++ b/test/stage1/behavior/new_stack_call.zig
@@ -8,11 +8,6 @@ test "calling a function with a new stack" {
     if (@import("builtin").arch == .aarch64) return error.SkipZigTest;
     if (@import("builtin").arch == .mipsel) return error.SkipZigTest;
 
-    if (@import("builtin").arch == .riscv64) {
-        // TODO: https://github.com/ziglang/zig/issues/3338
-        return error.SkipZigTest;
-    }
-
     const arg = 1234;
 
     const a = @newStackCall(new_stack_bytes[0..512], targetFunction, arg);
diff --git a/test/stage1/behavior/widening.zig b/test/stage1/behavior/widening.zig
index 1aa7ff5d..71f8ae48 100644
--- a/test/stage1/behavior/widening.zig
+++ b/test/stage1/behavior/widening.zig
@@ -19,10 +19,6 @@ test "implicit unsigned integer to signed integer" {
 }
 
 test "float widening" {
-    if (@import("builtin").arch == .riscv64) {
-        // TODO:
-        return error.SkipZigTest;
-    }
     var a: f16 = 12.34;
     var b: f32 = a;
     var c: f64 = b;
@@ -35,10 +31,6 @@ test "float widening" {
 test "float widening f16 to f128" {
     // TODO https://github.com/ziglang/zig/issues/3282
     if (@import("builtin").arch == .aarch64) return error.SkipZigTest;
-    if (@import("builtin").arch == .riscv64) {
-        // TODO: https://github.com/ziglang/zig/issues/3338
-        return error.SkipZigTest;
-    }
 
     var x: f16 = 12.34;
     var y: f128 = x;

2. Execute the following command (from the build directory):

./zig test ../test/stage1/behavior.zig -target riscv64-linux --test-cmd qemu-riscv64 --test-cmd-bin

The first problem you will run into will be:

LLVM ERROR: Named registers not implemented for this target

[luismarques]

I have the zig test suite passing after writing some fixes for these issues. Now I just need to transform the fixes into proper patches, with tests and so on. I'll see if those should be back-ported to LLVM 9.

[andrewrk]

That was fast! Let me know when this stuff makes it into trunk and I'll try a full test suite with building musl & glibc and everything, and try to find if there is anything else.

[luismarques]

The patches have been posted for review. When they both land I'll let you know.

[luismarques]

I believe this should now work in HEAD. Sorry for not indicating that immediately after the last fix landed.

[andrewrk]

Excellent! I'll leave this issue open until we update llvm dependency to 10.0.0, and enable the tests.